Reproduce “Dr. Post-Training: A Data Regularization Perspective on LLM Post-Training” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Pingbang Hu, Xueshen Liu, Z. Morley Mao, Jiaqi W. Ma

Dr. Post-Training lets target data guide model updates while abundant general data constrain them, selecting separate subsets for parameter groups, including individual layers. The best partition depends on mismatch and noise. On mismatched TyDiQA question-answering, layer-wise updates scored 48.3 ± 1.7 F1, versus 13.7 ± 1.5 globally and 4.9 ± 0.3 with standard training; results are not uniform across settings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.